Memory MCP is designed to solve the pain point that AI assistants start from scratch in every new session: users need to repeatedly explain the project architecture and paste the same patterns, causing the context window to be crowded out by duplicate content. Common memory solutions require a tool call for every lookup, adding latency and consuming Claude's thinking budget. Memory MCP solves this problem through a two-tier architecture.
[Main Capabilities]
- Hot cache (0ms): Through the SessionStart and UserPromptSubmit hooks, the plugin prints high-frequency knowledge directly into the conversation before Claude starts thinking, without any tool calls. The hook command only prints when the text differs from what was last seen in the current session; overhead is extremely low when the set is stable, and the whole set is reprinted when it changes.
- Cold storage (about 50ms): The remaining memories are searched by meaning using semantic similarity.
- Automatic promotion and demotion: Memories used more than 3 times are automatically promoted (about 20 items serve as the backing store for promoted memories), and memories unused for 14 days are automatically demoted. The hot cache holds about 10 items, composed of recently recalled memories, the predicted next memory, and the highest-priority promoted items.
- Knowledge graph: Extract knowledge from Claude's output and link it into the graph.
[How to Use]
- Add it to ~/.claude.json for configuration (see the official Reference for the full configuration options).
- Restart Claude Code.
- Run memory-mcp-cli bootstrap once to seed memories from project documentation; the hook only injects promoted content, and no content is seeded automatically.
- The first run automatically downloads an embedding model of about 90MB, taking about 30-60 seconds (only once).
- The plugin provides automatically configured hooks and slash commands (10 slash commands in total and a complete tool API; see the Reference).
- The dashboard lets you browse memories, the hot cache, injection history, sessions, and the knowledge graph.
[Applicable Scenarios] Suitable for Claude Code users who need to maintain project context across sessions and reduce repeated explanations, especially projects that frequently use fixed architecture patterns and code patterns.
[Limitations] The system consists of three complementary components (the plugin, the MCP server, and the dashboard), and most users are recommended to use the plugin; specific configuration details, commands, and APIs require consulting the official Reference, and this material does not provide a complete list. Pricing, cost, and platform details are not yet provided.
